Retrieval pipelines, vector search, chunking and reranking: how models are grounded in a corpus instead of their weights.
arXiv:2609. 03487v1 Announce Type: cross Abstract: Knowledge graph embedding (KGE) demonstrates its effectiveness for predicting missing links in knowledge graphs (KGs) by projecting entities and relations into a low-dimensional vector space.
Spruce is a system that enables scalable private outsourced retrieval by learning compact binary embeddings and using efficient Hamming-distance computation under a two‑server multi‑party computation protocol. It replaces costly corpus‑wide embedding scoring with a fixed‑radius protocol that avoids multi‑round candidate selection, and introduces private cluster pruning and a one‑core dealer to reduce computation and eliminate OT preprocessing bottlenecks. Across corpora of 383K–5.42M documents, Spruce maintains original search quality while achieving up to 31.5× higher throughput and reducing query times to a few seconds.
The paper introduces TEmBed, a unified benchmark for evaluating tabular embeddings across four representation levels—cell, row, column, and table—using a diverse set of models. It demonstrates that the best model depends on the specific task and representation level, providing practical guidance for selecting embeddings in real-world applications. The study aims to facilitate the development of more general-purpose tabular representation models.
The study maps tree species across Denmark using National Forest Inventory plots and Earth Observation data, comparing manually engineered spectral‑temporal features (STF) from Sentinel‑1 and Sentinel‑2 with embeddings from the foundation models TESSERA and AlphaEarth. Random forest, XGBoost, and MLP classifiers were evaluated, with the STF‑based MLP achieving the highest macro F1 scores for pure and mixed stands. The best model was applied nationally to produce a 10 m resolution tree species map, achieving 79.9% area‑adjusted accuracy and released as an open‑access product.
FinRAG-QA is a new benchmark dataset for financial question answering, featuring 999 practitioner-curated questions on 10 standardised indicators drawn from 209 annual and Pillar 3 reports of 24 major European and U.S. banks between 2019 and 2023. The dataset focuses on cross‑institutional retrieval over documents averaging 198k words, making it longer than any existing financial QA resource. Experiments on a multi‑stage Retrieval‑Augmented Generation pipeline show that contextual chunk enrichment and a retrieval‑optimised embedding model significantly improve NDCG@10, while a reasoning‑optimised generator boosts answer accuracy from 44.6% to 79.0% when the correct document is retrieved.
The paper introduces TRIM, a black‑box defense for backdoor attacks in computer vision models. TRIM identifies and removes malicious trigger regions at inference time using region‑based segmentation, adaptive trigger discovery via inpainting and diffusion, and selective purification, without needing model internals, training data, or clean samples. Experiments on various datasets and trigger types show TRIM reduces attack success rates to as low as 1.16% while maintaining high clean accuracy.
arXiv:2609. 04046v1 Announce Type: cross Abstract: What can a single layer of self-attention compute?
The Dice Roll Method is a standardized protocol for auditing large language model brand recommendations through repeated queries. It decomposes total response variance into sampling, prompt‑phrasing, run‑to‑run, and model‑version components, and uses a negative‑binomial mixed model, Cliff’s delta, and bootstrap techniques to guide iteration counts. The study identifies three iteration tiers—exploratory (n=5), confirmatory (n=10), and rigorous (n=15)—and recommends a compact battery of four complementary metrics for robust evaluation.
The paper introduces PACE, a dataset that tests whether personalized assistants can detect hidden conflicts in user requests by integrating implicit, egocentric knowledge from a knowledge base. PACE pairs user requests with persona-specific facts, requiring models to retrieve contextual evidence before deciding if a request is inappropriate. The authors also propose PaceMaker, a multi-agent system that improves conflict detection by coordinating query reformulation, multi-hop graph traversal, and conflict-aware filtering, outperforming existing methods on the PACE benchmark.
The paper introduces CRAW, a codec‑robust audio watermarking framework designed to embed imperceptible signals into synthetic speech. CRAW enhances robustness against neural re‑synthesis, codecs, denoisers, and vocoders while preserving high perceptual quality through distortion‑aware training, attention‑based pooling, perceptual masking, and error‑correcting codes. Experiments show CRAW outperforms existing post‑hoc watermarking methods in robustness without compromising audio quality.
Reflect‑SQL is a new framework for converting natural language into SQL queries. It tackles challenges such as large, obscure database schemas, poor table and column retrieval, and syntactically or logically flawed SQL by using a multi‑stage self‑reflection approach. The system iteratively refines queries and SQL through feedback loops driven by an LLM‑as‑a‑judge, achieving 72.03% execution accuracy on the BIRD benchmark, outperforming existing baselines.
PrivateHub is a contrastive diffusion model designed to generate synthetic multi‑sensor data that protects private user activities while keeping non‑private applications detectable. It operates in two stages: App‑Conditioned Pre‑training, which conditions the model on application embeddings, and App‑Aware Fine‑tuning, which uses contrastive learning to separate private from non‑private data. Experiments on three real‑world datasets demonstrate that PrivateHub reduces private‑application inference accuracy by 40–50% without harming non‑private performance and remains robust even when attackers retrain on the synthetic data.
The paper introduces a prompt‑engineering framework that personalizes large language model (LLM) teaching assistants across disciplines by tailoring responses to six learner‑specific dimensions, creating 96 distinct learner profiles. It also analyzes student queries through Bloom’s Taxonomy to gauge cognitive complexity, encoding both learner attributes and cognitive assessments into structured prompts that condition the LLM without retraining. Experiments using NLP metrics and a small human study demonstrate that this approach yields perceptible differences in response style and structure, with statistical evidence linking specific learner attributes to measurable changes.
The paper introduces Traceable TTS, a framework that enables Text‑to‑Speech systems to attribute synthesized speech to their source models without embedding explicit watermarks. By jointly training the TTS model and a discriminator, the method improves traceability generalization while maintaining or slightly enhancing audio quality. This represents the first attempt at watermark‑free TTS with strong traceability, and the authors plan to release the code to support further research.
The paper introduces STAIR, a retrieval system that uses a document’s Table of Contents to guide large language models in accessing global structure, thereby reducing hallucinations in Retrieval Augmented Generation. Experiments with a fine‑tuned Differentiable Search Index show that ToC‑based retrieval yields a low hallucination rate (<0.05%) and improves Recall@1 to 82.6% on the newly released SearchTome benchmark, outperforming baselines like BM25, DPR, and Mistral. The authors also release SearchTome, a diverse dataset of 18 books across six domains, to encourage further research in ToC‑based retrieval.
CORE improves compositional reasoning in multimodal language models by distilling a cross‑attentive reranker’s fine‑grained judgments into the embedding model. It generates candidate lists across five compositional matching levels and trains with a Rank‑KL objective to replicate the reranker’s ranking. Experiments on COLA, SUGARCREPE++, and NEGBENCH show CORE‑RERANKER‑8B outperforms Jina‑Reranker by 10.7 points, while CORE‑EMBED‑8B achieves the best overall average among evaluated embeddings, with gains also transferring to the MCMR benchmark without harming COCO or Flickr30K retrieval.
M3T introduces a discrete multi‑modal motion token system for sign language production, addressing the need for non‑manual features such as mouthings, eyebrow raises, gaze, and head movements. The approach couples FLAME’s expressive facial space with SMPL‑X body parameters and uses modality‑specific Finite Scalar Quantization VAEs to achieve high face codebook utilization (99.0%). Trained with an autoregressive transformer and a sign‑to‑text translation objective, M3T outperforms existing methods on three standard datasets, notably improving accuracy on NMFs‑CSL from 49.0% to 58.3% without large‑scale pre‑training.
The paper introduces OmniRSCLIP, an end‑to‑end contrastive learning framework that extends the CLIP architecture to handle heterogeneous remote sensing sensors such as SAR, multi‑spectral imaging, and hyperspectral imaging. It achieves this by employing Spectral‑Spatial Basis Decomposition to adapt arbitrary‑channel inputs without losing pretrained visual knowledge, and a spectral‑context‑aware mask‑based contrastive learning scheme to improve fine‑grained image‑text alignment. The authors also build OmniRS5M, a large‑scale image‑text corpus covering multiple sensor modalities, and demonstrate that OmniRSCLIP maintains strong RGB performance while effectively supporting these diverse remote sensing data types.
The paper introduces Alignment-Free Text‑Audiobox (Text‑AB), a unified diffusion‑based framework that performs high‑quality voice dubbing and full‑duplex dialogue synthesis without requiring forced alignment. Text‑AB uses a latent diffusion model with DAC‑VAE features, achieving over 10× compression compared to prior EnCodec representations, and learns text‑speech alignment via cross‑attention. The authors pretrain a 3B‑parameter model on 480k hours of monolingual speech and fine‑tune it for cross‑lingual dubbing, full‑duplex dialogue, and emotional dialogue, reporting significant improvements in prosody, voice similarity, naturalness, and emotional expressivity over existing internal systems.
The paper investigates using synthetic natural-language descriptions to contrastively pretrain small transformer encoders for code representation. By pairing generated descriptions with code in a dual-encoder setup during training and discarding them at inference, the authors achieve significant improvements over traditional pretraining baselines on most evaluated tasks. When fine‑tuned, these models match or surpass much larger zero‑shot models and remain competitive with execution‑aware supervision, indicating a scalable alternative for code embeddings.